Is your feature request related to a problem? Please describe.
The YOLO training manifest currently loads and stores every encoded video frame in memory, which can use significant memory for large datasets
Describe the solution you'd like
Load each image through the configured TokTagger data loader only when the PyTorch dataset requests it.
Describe alternatives you've considered
Continue storing encoded images in the manifest. This avoids repeated loading but uses more memory.
Additional context
As requested by #326 (comment)
Is your feature request related to a problem? Please describe.
The YOLO training manifest currently loads and stores every encoded video frame in memory, which can use significant memory for large datasets
Describe the solution you'd like
Load each image through the configured TokTagger data loader only when the PyTorch dataset requests it.
Describe alternatives you've considered
Continue storing encoded images in the manifest. This avoids repeated loading but uses more memory.
Additional context
As requested by #326 (comment)